TLC Network Group - Seminars

In the first part of the talk, I will provide an overview of ongoing research activities within the Networking Group at University of Thessaly.

In the second part of the talk, I will discuss the problem of distributed bandwidth sharing in peer-to-peer networks. In peer-to-peer networks, each peer plays the role of client and server. As server, it receives content requests of others and decides to what extent it will satisfy them by allocating upload bandwidth. As client, it sends its own requests to others to download content. We consider a star topology network with the capacity bottleneck being the peer access link to the backbone. Peers have different utility functions which are private information and capture a peer's selfishness or desire for content. Setting maximization of sum utility as the objective, we study bandwidth sharing between download flows of each peer and upload flows of others and how this can be performed in a decentralized autonomous fashion. We formulate and solve the problem using dual decomposition. En route to the solution, we devise meaningful reputation-driven protocol with the desirable property that only amounts of requested and granted bandwidth are circulated, and not reputations.

Speaker bio's: Iordanis Koutsopoulos is Lecturer (to be promoted to Assistant Professor) at the Department of Computer and Communications Engineering, University of Thessaly, Greece. He obtained the Diploma in Electrical and Computer Engineering from the National Technical University of Athens, Athens, Greece in 1997 and the M.S and Ph.D degrees in Electrical and Computer Engineering from the University of Maryland, College Park (UMCP) in 1999 and 2002 respectively. From 1997 to 2002 he was a Fulbright Fellow and a Research Assistant with the Institute for Systems Research (ISR) of UMCP. He has held internship positions with Hughes Network Systems, Germantown, MD, Hughes Research Laboratories LLC, Malibu, CA, and Aperto Networks Inc., Milpitas, CA, in 1998, 1999 and 2000 respectively. For the summer period of 2005 he was a visiting scholar with University of Washington, Seattle, WA. For the period 2005-2007 he was awarded the Marie Curie International Reintegration Grant (IRG). His research interests are in the field of networking with emphasis on wireless networks, cross-layer design, sensor networks, smart antennas and more recently on wireless network security and peer-to-peer networks.

Static multi-hop wireless (mesh) networks represent an emerging networking architecture which provides community networking, distributed sensing, support for medical applications, and Internet access in locations like airports, convention centers, education facilities, and hospitals. Mesh networks pose a challenge that does not exist in traditional wireline networks like the Internet: neighboring links interfere with each-other in complex ways. One way to address this issue is to design medium access schemes that carefully schedule the various competing links and mask the complex interference from higher layers. But, it is well know that such schedulers require a lot of computation and a centralized implementation. As a result, random access schedulers have become the de facto standard used in all deployments.

The first part of this talk addresses the following fundamental question: Is random access able to achieve efficient rates in the context of mess networks? The luck of a formal, comprehensive study of random access's performance has lead many researchers to believe that it performs significantly worse than optimal, collision-free schedulers. But this belief is either based on very specific topologies which cannot occur in practice due to physical layer limitations, or it is based on experiments that use the wrong transport protocols, e.g. TCP, for the task. Our work formally proves that random access scheduling is surprisingly close to optimal under realistic topologies. Specifically, we analytically compute the achievable rate region of the de facto standard random access scheme, 802.11, and find it very close to that of the optimal scheduler.

The second part of this talk addresses the following follow up practical question: Can we design lightweight, distributed rate controllers that realize this great rates? Our work shows via simulations as well as real world experiments that this is indeed the case. The main idea behind the design of our controllers lies in a paradigm shift; While for decades the networking community has been designing rate controllers by viewing a link (or node) as the point of potential congestion and regulating the rate of flows through that link (or node), we base our designs in a neighborhood-centric approach. In particular, when a node becomes congested, our schemes orchestrate the reduction of the rate of all flows which pass through the neighborhood of this node, rather than reducing the rate of only the flows which go through this node. Finally, we discuss practical ways to achieve the high performance of our clean-slate designs without changing a single line of code from the existing network stack.

Konstantinos Psounis is an assistant professor of Electrical Engineering and Computer Science at the University of Southern California. He received his first degree from the department of Electrical and Computer Engineering of National Technical University of Athens, Greece, in June 1997, the M.S. degree in Electrical Engineering from Stanford University, California, in January 1999, and the Ph.D. degree in Electrical Engineering from Stanford University in December 2002.

Konstantinos models and analyzes the performance of a variety of networks, including the Internet, mobile ad hoc networks, delay and disruptive tolerant networks, sensor networks, mesh networks, peer to peer networks and the web. He also designs methods and algorithms to solve problems related to such systems. He is the author of more than 50 research papers on these topics. Konstantinos has received faculty awards from the National Science Foundation, the Zumberge Foundation and Cisco Systems, has been elected a senior member at both IEEE and ACM, has been a Stanford graduate fellow throughout his graduate studies, and has received the best-student National Technical University of Athens award for graduating first in his class.

Static multi-hop wireless (mesh) networks represent an emerging networking architecture which provides community networking, distributed sensing, support for medical applications, and Internet access in locations like airports, convention centers, education facilities, and hospitals. Mesh networks pose a challenge that does not exist in traditional wireline networks like the Internet: neighboring links interfere with each-other in complex ways. One way to address this issue is to design medium access schemes that carefully schedule the various competing links and mask the complex interference from higher layers. But, it is well know that such schedulers require a lot of computation and a centralized implementation. As a result, random access schedulers have become the de facto standard used in all deployments.

The first part of this talk addresses the following fundamental question: Is random access able to achieve efficient rates in the context of mess networks? The luck of a formal, comprehensive study of random access's performance has lead many researchers to believe that it performs significantly worse than optimal, collision-free schedulers. But this belief is either based on very specific topologies which cannot occur in practice due to physical layer limitations, or it is based on experiments that use the wrong transport protocols, e.g. TCP, for the task. Our work formally proves that random access scheduling is surprisingly close to optimal under realistic topologies. Specifically, we analytically compute the achievable rate region of the de facto standard random access scheme, 802.11, and find it very close to that of the optimal scheduler.

The second part of this talk addresses the following follow up practical question: Can we design lightweight, distributed rate controllers that realize this great rates? Our work shows via simulations as well as real world experiments that this is indeed the case. The main idea behind the design of our controllers lies in a paradigm shift; While for decades the networking community has been designing rate controllers by viewing a link (or node) as the point of potential congestion and regulating the rate of flows through that link (or node), we base our designs in a neighborhood-centric approach. In particular, when a node becomes congested, our schemes orchestrate the reduction of the rate of all flows which pass through the neighborhood of this node, rather than reducing the rate of only the flows which go through this node. Finally, we discuss practical ways to achieve the high performance of our clean-slate designs without changing a single line of code from the existing network stack.

Konstantinos Psounis is an assistant professor of Electrical Engineering and Computer Science at the University of Southern California. He received his first degree from the department of Electrical and Computer Engineering of National Technical University of Athens, Greece, in June 1997, the M.S. degree in Electrical Engineering from Stanford University, California, in January 1999, and the Ph.D. degree in Electrical Engineering from Stanford University in December 2002.

Konstantinos models and analyzes the performance of a variety of networks, including the Internet, mobile ad hoc networks, delay and disruptive tolerant networks, sensor networks, mesh networks, peer to peer networks and the web. He also designs methods and algorithms to solve problems related to such systems. He is the author of more than 50 research papers on these topics. Konstantinos has received faculty awards from the National Science Foundation, the Zumberge Foundation and Cisco Systems, has been elected a senior member at both IEEE and ACM, has been a Stanford graduate fellow throughout his graduate studies, and has received the best-student National Technical University of Athens award for graduating first in his class.

Design and implementation of a new scheduling strategy for properly distributing the network available bandwidth among base and enhancement layers in quality adaptive video-streaming systems. The main target of the strategy is avoiding playout interruptions due to the underflows of the base layer buffer at the decoder side.

The energy consumption is becoming more and more a key issue, raising interest of people and of the research community to try to limit the energy waste. The Internet makes no exception, and several projects are studying how to reduce its energy consumption. In this paper, we provide a first evaluation of the the amount of redundant resources (nodes and links) that can be powered off from a network topology to reduce power consumption. We first formulate a theoretical evaluation that exploits random graph theory to estimate the fraction of nodes and links that can be removed from the topology without losing the connectivity property. Then we compare theoretical results with simulation results using realistic Internet topologies. Results, although preliminary, show that in the current Internet the amount of redundancy is large, so that up to 80% of transport nodes can be removed. Large energy saving can then be achieved by accurately turning off nodes and links, e.g., during off-peak time. We show also that the non-cooperative design of the current Internet severely impacts the possible energy saving, suggesting that a cooperative approach can be investigated further.

Research and development activities at CSP regarding: Software Defined Radio, Cognitive radio networks, Protocols for integrating DMR/Internet, and Personal HPCN.

Currently, all kinds of networks are evolving towards faster and more efficient data traffic transport architectures in response to the continuous growth of the internet. In the Metropolitan Area Network (MAN) segment, data traffic is growing at a rate that could soon lead to a saturation of the actual Sonet/SDH and Gigabit Ethernet equipment relying on the electronic processing of the whole node's incoming traffic. This kind of equipment requires a considerable amount of processing power which doesn't seem to grow as fast as the data traffic do.

Next generation MANs require more processing to be done in the optical domain in order to relieve the electronics from the burden of processing all the traffic passing through the node. In the past years, a lot of research effort has been devoted to the topic of next generation optical networks: one of the most promising architectures, both from the performance and feasibility point of view, uses the so called "broadcast and select over WDM" principle.

We consider a slotted input-queued switch with a crossbar-like switching fabric. In each time-slot, a centralized scheduler determines a switching fabric configuration by properly selecting input/output port connections. We consider the energy consumption needed to configure the switching fabric and we assume that the energy depends on the number of modifications in the switching configuration in two consecutive time-slots. We address the problem of scheduling a set of packets to minimize the required energy while preserving high throughput. We reduce the overall problem into the combination of two different optimization problems. We propose a family of algorithms to solve the problem and we discuss their energy-throughput performance.

Current research activity is focusing on introducing new services to a TDMA based MAC protocol named MS-ALOHA that allows a deterministic upper bound to channel access delay. that makes it suitable for QoS demanding applications. The second part of the presentation will give a brief introduction to possible future research activity in Home Area Networking.

Increasing attention has been devoted to software routers based on off-the-shelf hardware and open source operating systems running on Personal Computer (PC) architectures due to their flexibility and reduced cost.

However the software router with single PC has performance limitation because of centralized resources and lack scalability and resilience. To overcome those limitations, a multi-stage software router has been proposed. In multi-stage router, two or more off-the-shelf computers are coordinated to operate as a single router. This increases the number of router interfaces, and routing capability plus it has inherent redundancy. These benefits come at the cost of management complexity. The two or more devices inside the single router have to act as and work as a single networking device. For the purpose of this singleness, management software should coordinate the devices to extract unique information which represents the single router.

In this preliminary stage of the management plane of multi-stage software router research, we are considering the network management using SNMP and command line interface (CLI) to configure the device.

We extend the analysis of the scaling laws of wireless ad hoc networks to the case of correlated nodes movements, which are commonly found in real mobility processes. We consider a simple version of the Reference Point Group Mobility model, in which nodes belonging to the same group are constrained to lie in a disc area, whose center moves uniformly across the network according to the i.i.d. model. We assume fast mobility conditions, and take as primary goal the maximization of per-node throughput. We discover that correlated node movements have huge impact on asymptotic throughput and delay, and can sometimes lead to better performance than the one achievable under independent nodes movements.

The increasing success of P2P-TV applications, that may overwhelm the network with their large volume of traffic in the near future, calls for the need of new traffic models that can effectively represent the traffic generated by these applications. In this paper, we consider PPLive, one of the most popular P2P-TV applications today and we study and model the traffic generated by PPLive clients. From the analysis of some real traffic traces, we recognize that the peer may follow three typical behavior, depending on the peer degree of contribution to the video content distribution. We then propose two simple models of the data received or transmitted by a peer: i) the Memoryless model that fits the distribution of traffic exchanged during short time windows; ii) the Hidden-Markov model that introduces also some memory in the traffic generation process so that the autocorrelation function typical of real traces can be matched. The accuracy of the models for all three classes of peer behavior is validated by both directly comparing synthetic traces generated by the models with real traces and considering the performance of a queue fed by these traces. Our results show that the models are quite accurate and can be effectively used as synthetic traffic generators; the models can be helpful in many networking tasks such as network performance analysis, network planning and dimensioning, traffic engineering. From our results, it can be concluded that the considered P2PTV traffic is not that difficult to model, probably due to the fact that it is less closed-loop controlled and less adaptive to network conditions than other Internet traffic types.